In-Vehicle Perception Training Using Self-Supervised Image Signals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The development of Automated Driving Systems (ADS) faces challenges in continuously improving perception functionality due to the high cost and labor intensity of annotating large volumes of training data, as well as bandwidth and data privacy concerns, which hinder the efficient incorporation of new data and rare scenarios into machine-learning algorithms.
Innovation Solution
A self-supervised machine-learning algorithm generates outputs from ingested images, forming a supervisory signal for a supervised learning process to update the perception module's model parameters, enabling efficient and automated training without the need for annotated data, and facilitating the incorporation of diverse scenarios through decentralized federated learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If large volumes of annotated data are used to train machine-learning algorithms for ADS perception functions, then performance and reliability are improved, but cost and labor requirements increase significantly
Solution Approach 1:
The system uses self-supervised learning where the machine-learning algorithm generates its own supervisory signals from raw images without human annotation. The algorithm processes images through multiple layers to create feature representations that serve as training labels, enabling the system to improve itself automatically without external annotation resources.
Solution Approach 2:
The patent performs preliminary feature extraction and representation learning on raw images before the actual supervised training. By pre-processing images through self-supervised learning to generate feature maps and supervisory signals in advance, the system prepares training data that would otherwise require expensive human annotation, thus resolving the contradiction between performance and annotation cost.
2Reliability
If more training data including rare scenarios is collected to improve ADS reliability, then system robustness is enhanced, but bandwidth requirements and data privacy concerns increase
Solution Approach 1:
The patent extracts only the essential feature representations and supervisory signals from the raw image data through self-supervised learning, rather than storing and transmitting the complete annotated datasets. By taking out only the necessary training components (feature maps, loss functions, model updates), the system maintains robustness while minimizing data privacy risks and bandwidth requirements.
3Measurement precision
If traditional supervised learning with annotated data is used for training perception functions, then training accuracy is achieved, but training time and computational resources increase
Solution Approach 1:
The system performs preliminary self-supervised learning to generate feature representations and supervisory signals before the actual supervised training phase. This preliminary action prepares the training data in advance, allowing the supervised learning to converge faster with fewer iterations, thus reducing overall training time while maintaining accuracy.
Solution Approach 2:
The patent enables continuous learning by generating supervisory signals on-the-fly from raw images during normal operation. Instead of requiring batch processing of pre-annotated data, the system continuously extracts features and updates models in real-time, maintaining training accuracy while eliminating the time loss associated with manual annotation and batch processing.
Data Source
AI summary
A computer-implemented method for updating a perception function of a vehicle having an Automated Driving System (ADS) is disclosed. The ADS has a machine-learning algorithm for: generating an attention map or a feature map based on one or more ingested images and for providing one or more in-vehicle perception functions based on one or more ingested images. The method comprises obtaining one or more images of a scene in a surrounding environment of the vehicle, and updating one or more model parameters of the self-supervised machine-learning algorithm in accordance with a self-supervised machine learning process based on the obtained one or more images. The method further comprises generating a first output comprising an attention map or a feature map by processing the obtained one or more images by using the self-supervised machine-learning algorithm, and generating a supervisory signal for a supervised learning process based on the first output.